[RFC] [PATCH 0/2] Proposal to add EFI_MP_SERVICES_PROTOCOL support for AARCH64


Rebecca Cran
 

I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.

Patch 2/2 implements the EFI_MP_SERVICES_PROTOCOL for DXE in Library/MpInitLib.
CpuDxe initializes the MP support, and as a result gains a dependency on
MpInitLib. ArmVirt.dsc.inc needs updated to add the new library, as will
all AARCH64 platforms.

I sent out a patch for MdeModulePkg last week to add a corresponding test
application, MpServicesTest (see https://edk2.groups.io/g/devel/message/80878).

Rebecca Cran (2):
ArmPkg: Replace CoreId and ClusterId with Mpidr in ARM_CORE_INFO
struct
ArmPkg: Add Library/MpInitLib to support EFI_MP_SERVICES_PROTOCOL

ArmPkg/ArmPkg.dec | 4 +
ArmPkg/ArmPkg.dsc | 4 +
ArmPkg/Drivers/CpuDxe/AArch64/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/Arm/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/CpuDxe.c | 2 +
ArmPkg/Drivers/CpuDxe/CpuDxe.h | 10 +
ArmPkg/Drivers/CpuDxe/CpuDxe.inf | 6 +
ArmPkg/Drivers/CpuDxe/CpuMpInit.c | 604 ++++++++
ArmPkg/Include/Guid/ArmMpCoreInfo.h | 3 +-
ArmPkg/Include/Library/ArmLib.h | 4 +
ArmPkg/Include/Library/MpInitLib.h | 366 +++++
ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S | 61 +
ArmPkg/Library/MpInitLib/DxeMpInitLib.inf | 53 +
ArmPkg/Library/MpInitLib/DxeMpLib.c | 1498 ++++++++++++++++++++
ArmPkg/Library/MpInitLib/InternalMpInitLib.h | 358 +++++
ArmPlatformPkg/Library/ArmPlatformLibNull/ArmPlatformLibNull.c | 8 +-
ArmPlatformPkg/PrePeiCore/MainMPCore.c | 2 +-
ArmPlatformPkg/PrePi/MainMPCore.c | 2 +-
ArmVirtPkg/ArmVirt.dsc.inc | 3 +
19 files changed, 3024 insertions(+), 8 deletions(-)
create mode 100644 ArmPkg/Drivers/CpuDxe/AArch64/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/Arm/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/CpuMpInit.c
create mode 100644 ArmPkg/Include/Library/MpInitLib.h
create mode 100644 ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpInitLib.inf
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpLib.c
create mode 100644 ArmPkg/Library/MpInitLib/InternalMpInitLib.h

--
2.31.1


Leif Lindholm
 

On Fri, Sep 24, 2021 at 20:17:50 -0600, Rebecca Cran wrote:
I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.
Just to add to this:
Aff0 was never defined by the architecture to be the "core", it was
just the smallest schedulable entity. The intent being that whether
you had multiple hardware threads per core or not, you could just use
the affinity to determine whether
There is also a bit in the MPIDR to indicate whether the core *had*
multiple hardware threads.

In recent processors (without any change to the architecture), ARM
thought it would be beneficial to keep software developers on their
toes by starting to use the hyperthreading layout even for processors
without hyperthreading support. I.e. Aff0 is always 0 even though MT
is 0:
https://developer.arm.com/documentation/100798/0301/Register-descriptions/AArch64-system-registers/MPIDR-EL1--Multiprocessor-Affinity-Register--EL1
The justification being that an SoC might contain both processors
with and without multiple hardware threads per core.

Anyway, the point is that from at least Cortex-A76 onwards, Aff0 no
longer maps to CoreId universally, and Aff1 no longer maps to
ClusterId, for all non-threaded implementations.
So we need to start cleaning up this use.
This will possibly break some out-of-tree platforms, but I figure
we're far enough from next stable tag for that not to matter too
much.

Patch 2/2 implements the EFI_MP_SERVICES_PROTOCOL for DXE in Library/MpInitLib.
Worth calling out in the cover letter that this is backed by PSCI.

/
Leif

CpuDxe initializes the MP support, and as a result gains a dependency on
MpInitLib. ArmVirt.dsc.inc needs updated to add the new library, as will
all AARCH64 platforms.

I sent out a patch for MdeModulePkg last week to add a corresponding test
application, MpServicesTest (see https://edk2.groups.io/g/devel/message/80878).

Rebecca Cran (2):
ArmPkg: Replace CoreId and ClusterId with Mpidr in ARM_CORE_INFO
struct
ArmPkg: Add Library/MpInitLib to support EFI_MP_SERVICES_PROTOCOL

ArmPkg/ArmPkg.dec | 4 +
ArmPkg/ArmPkg.dsc | 4 +
ArmPkg/Drivers/CpuDxe/AArch64/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/Arm/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/CpuDxe.c | 2 +
ArmPkg/Drivers/CpuDxe/CpuDxe.h | 10 +
ArmPkg/Drivers/CpuDxe/CpuDxe.inf | 6 +
ArmPkg/Drivers/CpuDxe/CpuMpInit.c | 604 ++++++++
ArmPkg/Include/Guid/ArmMpCoreInfo.h | 3 +-
ArmPkg/Include/Library/ArmLib.h | 4 +
ArmPkg/Include/Library/MpInitLib.h | 366 +++++
ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S | 61 +
ArmPkg/Library/MpInitLib/DxeMpInitLib.inf | 53 +
ArmPkg/Library/MpInitLib/DxeMpLib.c | 1498 ++++++++++++++++++++
ArmPkg/Library/MpInitLib/InternalMpInitLib.h | 358 +++++
ArmPlatformPkg/Library/ArmPlatformLibNull/ArmPlatformLibNull.c | 8 +-
ArmPlatformPkg/PrePeiCore/MainMPCore.c | 2 +-
ArmPlatformPkg/PrePi/MainMPCore.c | 2 +-
ArmVirtPkg/ArmVirt.dsc.inc | 3 +
19 files changed, 3024 insertions(+), 8 deletions(-)
create mode 100644 ArmPkg/Drivers/CpuDxe/AArch64/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/Arm/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/CpuMpInit.c
create mode 100644 ArmPkg/Include/Library/MpInitLib.h
create mode 100644 ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpInitLib.inf
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpLib.c
create mode 100644 ArmPkg/Library/MpInitLib/InternalMpInitLib.h

--
2.31.1


Leif Lindholm
 

Ard/Sami - any comments?

/
Leif

On Tue, Sep 28, 2021 at 12:14:35 +0100, Leif Lindholm wrote:
On Fri, Sep 24, 2021 at 20:17:50 -0600, Rebecca Cran wrote:
I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.
Just to add to this:
Aff0 was never defined by the architecture to be the "core", it was
just the smallest schedulable entity. The intent being that whether
you had multiple hardware threads per core or not, you could just use
the affinity to determine whether
There is also a bit in the MPIDR to indicate whether the core *had*
multiple hardware threads.

In recent processors (without any change to the architecture), ARM
thought it would be beneficial to keep software developers on their
toes by starting to use the hyperthreading layout even for processors
without hyperthreading support. I.e. Aff0 is always 0 even though MT
is 0:
https://developer.arm.com/documentation/100798/0301/Register-descriptions/AArch64-system-registers/MPIDR-EL1--Multiprocessor-Affinity-Register--EL1
The justification being that an SoC might contain both processors
with and without multiple hardware threads per core.

Anyway, the point is that from at least Cortex-A76 onwards, Aff0 no
longer maps to CoreId universally, and Aff1 no longer maps to
ClusterId, for all non-threaded implementations.
So we need to start cleaning up this use.
This will possibly break some out-of-tree platforms, but I figure
we're far enough from next stable tag for that not to matter too
much.

Patch 2/2 implements the EFI_MP_SERVICES_PROTOCOL for DXE in Library/MpInitLib.
Worth calling out in the cover letter that this is backed by PSCI.

/
Leif

CpuDxe initializes the MP support, and as a result gains a dependency on
MpInitLib. ArmVirt.dsc.inc needs updated to add the new library, as will
all AARCH64 platforms.

I sent out a patch for MdeModulePkg last week to add a corresponding test
application, MpServicesTest (see https://edk2.groups.io/g/devel/message/80878).

Rebecca Cran (2):
ArmPkg: Replace CoreId and ClusterId with Mpidr in ARM_CORE_INFO
struct
ArmPkg: Add Library/MpInitLib to support EFI_MP_SERVICES_PROTOCOL

ArmPkg/ArmPkg.dec | 4 +
ArmPkg/ArmPkg.dsc | 4 +
ArmPkg/Drivers/CpuDxe/AArch64/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/Arm/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/CpuDxe.c | 2 +
ArmPkg/Drivers/CpuDxe/CpuDxe.h | 10 +
ArmPkg/Drivers/CpuDxe/CpuDxe.inf | 6 +
ArmPkg/Drivers/CpuDxe/CpuMpInit.c | 604 ++++++++
ArmPkg/Include/Guid/ArmMpCoreInfo.h | 3 +-
ArmPkg/Include/Library/ArmLib.h | 4 +
ArmPkg/Include/Library/MpInitLib.h | 366 +++++
ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S | 61 +
ArmPkg/Library/MpInitLib/DxeMpInitLib.inf | 53 +
ArmPkg/Library/MpInitLib/DxeMpLib.c | 1498 ++++++++++++++++++++
ArmPkg/Library/MpInitLib/InternalMpInitLib.h | 358 +++++
ArmPlatformPkg/Library/ArmPlatformLibNull/ArmPlatformLibNull.c | 8 +-
ArmPlatformPkg/PrePeiCore/MainMPCore.c | 2 +-
ArmPlatformPkg/PrePi/MainMPCore.c | 2 +-
ArmVirtPkg/ArmVirt.dsc.inc | 3 +
19 files changed, 3024 insertions(+), 8 deletions(-)
create mode 100644 ArmPkg/Drivers/CpuDxe/AArch64/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/Arm/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/CpuMpInit.c
create mode 100644 ArmPkg/Include/Library/MpInitLib.h
create mode 100644 ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpInitLib.inf
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpLib.c
create mode 100644 ArmPkg/Library/MpInitLib/InternalMpInitLib.h

--
2.31.1


Ard Biesheuvel
 

On Thu, 7 Oct 2021 at 11:41, Leif Lindholm <leif@nuviainc.com> wrote:

Ard/Sami - any comments?
The code changes by itself look fine to me. The only problem I see is
that we cannot run arbitrary code on other cores with the MMU off,
given that in that case, they don't comply with the UEFI AArch64
platform requirements, most notably the one that says that unaligned
accesses must be permitted.

So either we severely constrain the kind of code that we permit to run
on other cores, or we enable the MMU and caches on each core as it
comes out of reset, as well as do any other CPU specific
initialization that we do for the primary core as well.



On Tue, Sep 28, 2021 at 12:14:35 +0100, Leif Lindholm wrote:
On Fri, Sep 24, 2021 at 20:17:50 -0600, Rebecca Cran wrote:
I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.
Just to add to this:
Aff0 was never defined by the architecture to be the "core", it was
just the smallest schedulable entity. The intent being that whether
you had multiple hardware threads per core or not, you could just use
the affinity to determine whether
There is also a bit in the MPIDR to indicate whether the core *had*
multiple hardware threads.

In recent processors (without any change to the architecture), ARM
thought it would be beneficial to keep software developers on their
toes by starting to use the hyperthreading layout even for processors
without hyperthreading support. I.e. Aff0 is always 0 even though MT
is 0:
https://developer.arm.com/documentation/100798/0301/Register-descriptions/AArch64-system-registers/MPIDR-EL1--Multiprocessor-Affinity-Register--EL1
The justification being that an SoC might contain both processors
with and without multiple hardware threads per core.

Anyway, the point is that from at least Cortex-A76 onwards, Aff0 no
longer maps to CoreId universally, and Aff1 no longer maps to
ClusterId, for all non-threaded implementations.
So we need to start cleaning up this use.
This will possibly break some out-of-tree platforms, but I figure
we're far enough from next stable tag for that not to matter too
much.

Patch 2/2 implements the EFI_MP_SERVICES_PROTOCOL for DXE in Library/MpInitLib.
Worth calling out in the cover letter that this is backed by PSCI.

/
Leif

CpuDxe initializes the MP support, and as a result gains a dependency on
MpInitLib. ArmVirt.dsc.inc needs updated to add the new library, as will
all AARCH64 platforms.

I sent out a patch for MdeModulePkg last week to add a corresponding test
application, MpServicesTest (see https://edk2.groups.io/g/devel/message/80878).

Rebecca Cran (2):
ArmPkg: Replace CoreId and ClusterId with Mpidr in ARM_CORE_INFO
struct
ArmPkg: Add Library/MpInitLib to support EFI_MP_SERVICES_PROTOCOL

ArmPkg/ArmPkg.dec | 4 +
ArmPkg/ArmPkg.dsc | 4 +
ArmPkg/Drivers/CpuDxe/AArch64/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/Arm/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/CpuDxe.c | 2 +
ArmPkg/Drivers/CpuDxe/CpuDxe.h | 10 +
ArmPkg/Drivers/CpuDxe/CpuDxe.inf | 6 +
ArmPkg/Drivers/CpuDxe/CpuMpInit.c | 604 ++++++++
ArmPkg/Include/Guid/ArmMpCoreInfo.h | 3 +-
ArmPkg/Include/Library/ArmLib.h | 4 +
ArmPkg/Include/Library/MpInitLib.h | 366 +++++
ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S | 61 +
ArmPkg/Library/MpInitLib/DxeMpInitLib.inf | 53 +
ArmPkg/Library/MpInitLib/DxeMpLib.c | 1498 ++++++++++++++++++++
ArmPkg/Library/MpInitLib/InternalMpInitLib.h | 358 +++++
ArmPlatformPkg/Library/ArmPlatformLibNull/ArmPlatformLibNull.c | 8 +-
ArmPlatformPkg/PrePeiCore/MainMPCore.c | 2 +-
ArmPlatformPkg/PrePi/MainMPCore.c | 2 +-
ArmVirtPkg/ArmVirt.dsc.inc | 3 +
19 files changed, 3024 insertions(+), 8 deletions(-)
create mode 100644 ArmPkg/Drivers/CpuDxe/AArch64/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/Arm/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/CpuMpInit.c
create mode 100644 ArmPkg/Include/Library/MpInitLib.h
create mode 100644 ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpInitLib.inf
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpLib.c
create mode 100644 ArmPkg/Library/MpInitLib/InternalMpInitLib.h

--
2.31.1


Leif Lindholm
 

On Thu, Oct 07, 2021 at 12:03:30 +0200, Ard Biesheuvel wrote:
On Thu, 7 Oct 2021 at 11:41, Leif Lindholm <leif@nuviainc.com> wrote:

Ard/Sami - any comments?
The code changes by itself look fine to me. The only problem I see is
that we cannot run arbitrary code on other cores with the MMU off,
given that in that case,
This is a good point, and something we at the very least should point
out more explicitly - and perhaps ought to provide a helper library to
do - ...

they don't comply with the UEFI AArch64
platform requirements, most notably the one that says that unaligned
accesses must be permitted.
... but I'll note that EFI_MP_SERVICES_PROTOCOL is one of those
incorrectly named protocols defined in PI, not UEFI.

So either we severely constrain the kind of code that we permit to run
on other cores, or we enable the MMU and caches on each core as it
comes out of reset, as well as do any other CPU specific
initialization that we do for the primary core as well.
The description for StartupAllAPs() has a note:
It is the responsibility of the consumer of the
EFI_MP_SERVICES_PROTOCOL.StartupAllAPs() to make sure that the nature
of the code that is executed on the BSP and the dispatched APs is well
controlled. The MP Services Protocol does not guarantee that the
Procedure function is MP-safe. Hence, the tasks that can be run in
parallel are limited to certain independent tasks and well-controlled
exclusive code. EFI services and protocols may not be called by APs
unless otherwise specified.

So I think this is actually fine, implementation-wise. *Except* for
the SwitchBSP function (where we're currently bailing out anyway).

Regards,

Leif

On Tue, Sep 28, 2021 at 12:14:35 +0100, Leif Lindholm wrote:
On Fri, Sep 24, 2021 at 20:17:50 -0600, Rebecca Cran wrote:
I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.
Just to add to this:
Aff0 was never defined by the architecture to be the "core", it was
just the smallest schedulable entity. The intent being that whether
you had multiple hardware threads per core or not, you could just use
the affinity to determine whether
There is also a bit in the MPIDR to indicate whether the core *had*
multiple hardware threads.

In recent processors (without any change to the architecture), ARM
thought it would be beneficial to keep software developers on their
toes by starting to use the hyperthreading layout even for processors
without hyperthreading support. I.e. Aff0 is always 0 even though MT
is 0:
https://developer.arm.com/documentation/100798/0301/Register-descriptions/AArch64-system-registers/MPIDR-EL1--Multiprocessor-Affinity-Register--EL1
The justification being that an SoC might contain both processors
with and without multiple hardware threads per core.

Anyway, the point is that from at least Cortex-A76 onwards, Aff0 no
longer maps to CoreId universally, and Aff1 no longer maps to
ClusterId, for all non-threaded implementations.
So we need to start cleaning up this use.
This will possibly break some out-of-tree platforms, but I figure
we're far enough from next stable tag for that not to matter too
much.

Patch 2/2 implements the EFI_MP_SERVICES_PROTOCOL for DXE in Library/MpInitLib.
Worth calling out in the cover letter that this is backed by PSCI.

/
Leif

CpuDxe initializes the MP support, and as a result gains a dependency on
MpInitLib. ArmVirt.dsc.inc needs updated to add the new library, as will
all AARCH64 platforms.

I sent out a patch for MdeModulePkg last week to add a corresponding test
application, MpServicesTest (see https://edk2.groups.io/g/devel/message/80878).

Rebecca Cran (2):
ArmPkg: Replace CoreId and ClusterId with Mpidr in ARM_CORE_INFO
struct
ArmPkg: Add Library/MpInitLib to support EFI_MP_SERVICES_PROTOCOL

ArmPkg/ArmPkg.dec | 4 +
ArmPkg/ArmPkg.dsc | 4 +
ArmPkg/Drivers/CpuDxe/AArch64/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/Arm/Arch.c | 22 +
ArmPkg/Drivers/CpuDxe/CpuDxe.c | 2 +
ArmPkg/Drivers/CpuDxe/CpuDxe.h | 10 +
ArmPkg/Drivers/CpuDxe/CpuDxe.inf | 6 +
ArmPkg/Drivers/CpuDxe/CpuMpInit.c | 604 ++++++++
ArmPkg/Include/Guid/ArmMpCoreInfo.h | 3 +-
ArmPkg/Include/Library/ArmLib.h | 4 +
ArmPkg/Include/Library/MpInitLib.h | 366 +++++
ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S | 61 +
ArmPkg/Library/MpInitLib/DxeMpInitLib.inf | 53 +
ArmPkg/Library/MpInitLib/DxeMpLib.c | 1498 ++++++++++++++++++++
ArmPkg/Library/MpInitLib/InternalMpInitLib.h | 358 +++++
ArmPlatformPkg/Library/ArmPlatformLibNull/ArmPlatformLibNull.c | 8 +-
ArmPlatformPkg/PrePeiCore/MainMPCore.c | 2 +-
ArmPlatformPkg/PrePi/MainMPCore.c | 2 +-
ArmVirtPkg/ArmVirt.dsc.inc | 3 +
19 files changed, 3024 insertions(+), 8 deletions(-)
create mode 100644 ArmPkg/Drivers/CpuDxe/AArch64/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/Arm/Arch.c
create mode 100644 ArmPkg/Drivers/CpuDxe/CpuMpInit.c
create mode 100644 ArmPkg/Include/Library/MpInitLib.h
create mode 100644 ArmPkg/Library/MpInitLib/AArch64/MpFuncs.S
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpInitLib.inf
create mode 100644 ArmPkg/Library/MpInitLib/DxeMpLib.c
create mode 100644 ArmPkg/Library/MpInitLib/InternalMpInitLib.h

--
2.31.1


Ard Biesheuvel
 

On Thu, 7 Oct 2021 at 13:02, Leif Lindholm <leif@nuviainc.com> wrote:

On Thu, Oct 07, 2021 at 12:03:30 +0200, Ard Biesheuvel wrote:
On Thu, 7 Oct 2021 at 11:41, Leif Lindholm <leif@nuviainc.com> wrote:

Ard/Sami - any comments?
The code changes by itself look fine to me. The only problem I see is
that we cannot run arbitrary code on other cores with the MMU off,
given that in that case,
This is a good point, and something we at the very least should point
out more explicitly - and perhaps ought to provide a helper library to
do - ...

they don't comply with the UEFI AArch64
platform requirements, most notably the one that says that unaligned
accesses must be permitted.
... but I'll note that EFI_MP_SERVICES_PROTOCOL is one of those
incorrectly named protocols defined in PI, not UEFI.

So either we severely constrain the kind of code that we permit to run
on other cores, or we enable the MMU and caches on each core as it
comes out of reset, as well as do any other CPU specific
initialization that we do for the primary core as well.
The description for StartupAllAPs() has a note:
It is the responsibility of the consumer of the
EFI_MP_SERVICES_PROTOCOL.StartupAllAPs() to make sure that the nature
of the code that is executed on the BSP and the dispatched APs is well
controlled. The MP Services Protocol does not guarantee that the
Procedure function is MP-safe. Hence, the tasks that can be run in
parallel are limited to certain independent tasks and well-controlled
exclusive code. EFI services and protocols may not be called by APs
unless otherwise specified.

So I think this is actually fine, implementation-wise. *Except* for
the SwitchBSP function (where we're currently bailing out anyway).
Ok, so that doesn't look as bad as I thought. But we'll have to be
more strict than other arches: even EFI services and protocols that
are marked as safe for execution under this MP protocol are likely to
explode if they rely on CopyMem() or SetMem() for in/outputs that are
not a multiple of 8 bytes in case the platform uses the
BaseMemoryLibOptDxe flavour of this library, since it relies heavily
on deliberately misaligned loads and stores.


Leif Lindholm
 

+Samer

On Fri, Oct 8, 2021 at 3:51 PM Ard Biesheuvel <ardb@...> wrote:
> > So either we severely constrain the kind of code that we permit to run
> > on other cores, or we enable the MMU and caches on each core as it
> > comes out of reset, as well as do any other CPU specific
> > initialization that we do for the primary core as well.
>
> The description for StartupAllAPs() has a note:
> It is the responsibility of the consumer of the
> EFI_MP_SERVICES_PROTOCOL.StartupAllAPs() to make sure that the nature
> of the code that is executed on the BSP and the dispatched APs is well
> controlled. The MP Services Protocol does not guarantee that the
> Procedure function is MP-safe. Hence, the tasks that can be run in
> parallel are limited to certain independent tasks and well-controlled
> exclusive code. EFI services and protocols may not be called by APs
> unless otherwise specified.
>
> So I think this is actually fine, implementation-wise. *Except* for
> the SwitchBSP function (where we're currently bailing out anyway).

Ok, so that doesn't look as bad as I thought. But we'll have to be
more strict than other arches: even EFI services and protocols that
are marked as safe for execution under this MP protocol are likely to
explode if they rely on CopyMem() or SetMem() for in/outputs that are
not a multiple of 8 bytes in case the platform uses the
BaseMemoryLibOptDxe flavour of this library, since it relies heavily
on deliberately misaligned loads and stores.

I think there is no way a protocol defined in the UEFI specification could be
safe to use by non-BSP. In PI, the only references I find to the protocol are
in MM and SAL protocols.
And we're not even looking at EFI_MP_SERVICES_PPI at this point.

But it might be good to hear something from ARM whether the use of this
protocol which "must be produced on any system with more than one logical processor"
*should* be able to rely on anything being set up for it, or whether we
need an aforementioned helper library.

/
    Leif


Samer El-Haj-Mahmoud
 

> In PI, the only references I find to the protocol are in MM and SAL protocols.

> And we're not even looking at EFI_MP_SERVICES_PPI at this point.

 

The PI 1.7 spec defined the EFI_MP_SERVICES_PROTOCOL in page 2-180, with the PPI and MM versions in 1-193 and 4-57 respectively.

 

 

> But it might be good to hear something from ARM whether the use of this

> protocol which "must be produced on any system with more than one logical processor"

> *should* be able to rely on anything being set up for it, or whether we

> need an aforementioned helper library.

 

This statement (from the PI spec) is overly ambitious. I bet that it does not hold true today on most DXE-based UEFI implementations on other architectures, not just AARCH64. If we agree, I will file an ECR to remove this statement from the PI spec.

 

From AARCH64 SBBR systems point of view:

  • The requirements from Arm SBBR point of view are around using PSCI to online/offline Secondary cores, and leaving them offlined before ReadyToBoot is signaled.
  • PI-based UEFI implementations are not required. And even when they are implemented, the EFI_MP_SERVICES_PROTOCOL is not required
  • I agree with the analysis in this thread. EFI_MP implementations on AARCh64 need to be severely limited in the general case. Platforms (upstream or downstream) can still innovate and write their own code to run in these services as they wish.

 

 

Thanks,

--Samer

 

 

 

From: Leif Lindholm <leif@...>
Sent: Monday, October 11, 2021 8:28 AM
To: Ard Biesheuvel <ardb@...>; Samer El-Haj-Mahmoud <Samer.El-Haj-Mahmoud@...>
Cc: Ard Biesheuvel <ardb+tianocore@...>; Sami Mujawar <Sami.Mujawar@...>; edk2-devel-groups-io <devel@edk2.groups.io>; Rebecca Cran <rebecca@...>; Gerd Hoffmann <kraxel@...>; edk2 RFC list <rfc@edk2.groups.io>
Subject: Re: [RFC] [PATCH 0/2] Proposal to add EFI_MP_SERVICES_PROTOCOL support for AARCH64

 

+Samer

 

On Fri, Oct 8, 2021 at 3:51 PM Ard Biesheuvel <ardb@...> wrote:

> > So either we severely constrain the kind of code that we permit to run
> > on other cores, or we enable the MMU and caches on each core as it
> > comes out of reset, as well as do any other CPU specific
> > initialization that we do for the primary core as well.
>
> The description for StartupAllAPs() has a note:
> It is the responsibility of the consumer of the
> EFI_MP_SERVICES_PROTOCOL.StartupAllAPs() to make sure that the nature
> of the code that is executed on the BSP and the dispatched APs is well
> controlled. The MP Services Protocol does not guarantee that the
> Procedure function is MP-safe. Hence, the tasks that can be run in
> parallel are limited to certain independent tasks and well-controlled
> exclusive code. EFI services and protocols may not be called by APs
> unless otherwise specified.
>
> So I think this is actually fine, implementation-wise. *Except* for
> the SwitchBSP function (where we're currently bailing out anyway).

Ok, so that doesn't look as bad as I thought. But we'll have to be
more strict than other arches: even EFI services and protocols that
are marked as safe for execution under this MP protocol are likely to
explode if they rely on CopyMem() or SetMem() for in/outputs that are
not a multiple of 8 bytes in case the platform uses the
BaseMemoryLibOptDxe flavour of this library, since it relies heavily
on deliberately misaligned loads and stores.

 

I think there is no way a protocol defined in the UEFI specification could be

safe to use by non-BSP. In PI, the only references I find to the protocol are

in MM and SAL protocols.

And we're not even looking at EFI_MP_SERVICES_PPI at this point.

 

But it might be good to hear something from ARM whether the use of this

protocol which "must be produced on any system with more than one logical processor"

*should* be able to rely on anything being set up for it, or whether we

need an aforementioned helper library.

 

/

    Leif

 

IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.


Rebecca Cran
 

On 9/28/21 5:14 AM, Leif Lindholm wrote:

On Fri, Sep 24, 2021 at 20:17:50 -0600, Rebecca Cran wrote:
I'd like to propose adding EFI_MP_SERVICES_PROTOCOL support for
AARCH64 systems. I've attached two patches to implement support for it
in the DXE phase, based on code in EmulatorPkg and UefiCpuPkg. It's added
under ArmPkg for now, but longer term it should probably be moved into
UefiCpuPkg.

Patch 1/2 is the start of addressing the issue that the Aff0 field of
the MPIDR is no longer guaranteed to be the core, and should be referred
to in a more generic way: for example it could be the thread, with Aff1
being the core and Aff2 the cluster. Clearly much more work is needed
to fully remove that assumption.
Just to add to this:
Aff0 was never defined by the architecture to be the "core", it was
just the smallest schedulable entity. The intent being that whether
you had multiple hardware threads per core or not, you could just use
the affinity to determine whether
There is also a bit in the MPIDR to indicate whether the core *had*
multiple hardware threads.

In recent processors (without any change to the architecture), ARM
thought it would be beneficial to keep software developers on their
toes by starting to use the hyperthreading layout even for processors
without hyperthreading support. I.e. Aff0 is always 0 even though MT
is 0:
https://developer.arm.com/documentation/100798/0301/Register-descriptions/AArch64-system-registers/MPIDR-EL1--Multiprocessor-Affinity-Register--EL1
The justification being that an SoC might contain both processors
with and without multiple hardware threads per core.

Anyway, the point is that from at least Cortex-A76 onwards, Aff0 no
longer maps to CoreId universally, and Aff1 no longer maps to
ClusterId, for all non-threaded implementations.
So we need to start cleaning up this use.
This will possibly break some out-of-tree platforms, but I figure
we're far enough from next stable tag for that not to matter too
much.
This patch will also break out-of-tree platforms because it causes ArmPkg/Drivers/CpuDxe to gain a dependency on MpInitLib.


--
Rebecca Cran